screenpipe

screenpipe is a local-first memory layer: a desktop agent that records what you see and hear on your own machine, stores it locally, and makes it searchable by the AI tools you already use. Negentropy Labs, Inc., trading as screenpipe, ships it for macOS, Windows and Linux, and describes the product as source-available rather than open source. Capture works through the operating system rather than through integrations. The app reads on-screen text primarily via accessibility APIs with OCR as a fallback, transcribes audio locally in what its FAQ puts at about 100 languages, and writes everything to a local store on your own disk. The pricing page puts the cost of that at roughly 30 GB per month of continuous recording, while the About page FAQ still says 5 to 10 GB, and retention is configurable either way. On macOS the app needs both screen recording and accessibility access granted in System Settings before it captures anything. For retrieval, screenpipe runs a local MCP server, installed with npx -y screenpipe-mcp, and exposes exactly two documented tools: search-content, which queries recorded screen text, audio transcriptions and input events with content-type, time-range and pagination parameters, and export-video. The vendor documents one-click installation into Claude Desktop from the app's own settings panel, plus manual stdio configuration for Claude Code, OpenAI Codex, Cursor and Warp, and generic instructions for other editors. An optional HTTP transport covers clients that cannot speak stdio. Under all of it sits a local REST API on port 3030, which is also what the MCP server talks to, so a scheduled task or a script can read the same history without going through an AI client at all. Scheduled tasks are the automation half: the docs call them scheduled AI agents, and the CLI and local API still call them pipes. Each is a markdown file carrying a prompt and a schedule: every 30 minutes, every 2 hours, daily, a cron expression, or manual, which is the default until one is set. A running task can query the capture database, write files, call external APIs and send notifications. Permission presets restrict what a task can do through the screenpipe API, but the docs state that they do not sandbox shell commands, files or external tools. The documented connections list more than 50 apps, including Slack, Notion, Google Calendar, Obsidian, HubSpot, Salesforce, Zoom, PostHog and Sentry. The vendor's own examples are tracking time in Toggl from app usage, writing daily summaries to Obsidian, and sending a Slack message after 30 minutes on a site. Scheduled workflows are a priced line on the plan comparison, limited on Free and unlimited from Basic upward. Several of the vendor's published recipes are deliberately review-gated instead, drafting a follow-up queue or CRM fields for a person to approve. Pricing is published on the page the /pricing path resolves to, and there are four plans. Free is $0 on one device. Basic is $25 per month, or $21 per month billed annually at $250 per year. Business is $50 per seat per month, or $500 per seat per year. The seven-day onboarding trial requires a payment method, can be canceled before renewal, and then continues on the free tier. Enterprise is priced through a contact form, and it is the only tier with SSO and MDM controls. Students, researchers and faculty get 50% off the first month. Each plan includes a hosted AI allowance, and the vendor's documentation states that it does not fix a token quantity or reset schedule because these can vary by account. Background scheduled tasks draw on the same allowance as chat, there is no user-defined spending cap, and allowance warnings are advisory rather than guaranteed before every run. A user's own provider key can be used instead where supported. The About page still lists an older ladder of $25, $50 and $150 per month and its structured data still carries a $25 offer, so the two vendor surfaces differ. Currency is USD on both. On privacy the architecture and the legal terms address different things. Data stays on the device by default, cloud model calls send only the query, and Ollama can be used for fully local inference. The terms state that customer content, output and personal information are not used to train models without separate written agreement. The same section then has the user consent, by using the service, to the creation and use of Deidentified Data and Analytics Data including to improve and train the vendor's own models. The public repository carries 21.8k GitHub stars, and its license is not an OSI license: personal, non-commercial and evaluation use can be free while commercial use requires a paid license under the Screenpipe Commercial License. screenpipe suits individuals and small teams who want their own screen, meeting and app history captured on their own machine and queryable by the AI clients they already run. Teams that want meeting notes alone, without continuous screen capture, should look at Granola. Developers who want a memory layer to build into their own agent under an OSI open-source license should look at Mem0. Teams whose priority is searching the documentation and tickets already held in their existing tools, rather than their own captured activity, should look at Atlassian Rovo. Current state Q4 2026: the local MCP server documents two tools, search-content and export-video. The /pricing path resolves to a plan page listing Free, Basic, Business and Enterprise, while the About page carries an older three-tier ladder. GitHub reports the repository license as Other rather than an OSI license, and the Terms of Service, last updated 2 September 2026, name Negentropy Labs, Inc. as the operator.